Design Collaborative Document Editing

Keeping concurrent edits consistent

Several users edit the same document in real time and must converge on the same content. Clients apply edits locally for instant feedback and send operations (insert or delete at a position) over a WebSocket to a collaboration service. Two families of algorithms make concurrent edits converge. Operational transformation (OT) transforms an incoming operation against concurrent ones (if someone inserted earlier in the text, shift my position); it is typically run with a central server that orders operations. CRDTs (conflict-free replicated data types) give each character a unique, ordered identifier so replicas can merge in any order and still converge, which suits peer-to-peer and offline use at the cost of extra metadata. Around that: document sessions pinned to one server per document, an operation log plus periodic snapshots, presence (cursors) as ephemeral data, and version history.

Collaboration, scheduling and observability

These problems test real-time sync, reliable background work and high-volume data pipelines.

Why transformation is needed

Two concurrent inserts on "abc".

Start: "abc"
Alice: insert("X", pos=0)     -> "Xabc"
Bob:   insert("Y", pos=3)     -> "abcY"   (concurrent, based on "abc")

Without transformation, Alice applies Bob op at pos 3:
  "Xabc" -> "XabYc"   (wrong)
With OT, Bob op is transformed past Alice insert at pos 0 -> pos 4:
  "Xabc" -> "XabcY"   (both converge)

Architecture
clients <-> WebSocket gateway <-> doc session (one owner per doc id)
          session: order ops, transform/merge, append to op log
          storage: op log + snapshot every N ops; presence in memory

Directions that adjust

If a friend inserts a new street before yours on a shared map, "third left" becomes "fourth left". OT rewrites your instruction to account for their change.

Quick check: Which property makes CRDTs attractive for offline editing?

• They forbid concurrent edits
• They require a single central lock
• They never store metadata
• Replicas can merge operations in any order and still converge